Nvidia’s research argues that an AI harness (architecture managing context, memory and actions) can influence performance more than the underlying model. Its technical blog post, titled "Six Agent Harness Capabilities for Higher Model Performance," says harness design alone can deliver double-digit percentage improvements in benchmark scores while sharply reducing token consumption. Nvidia Labs also released NOOA, short for NVIDIA Labs Object-Oriented Agents, an open-source Python framework that structures AI agents as individual classes. The framework uses typed input and output, pass-by-reference memory management, code-based actions and model-callable APIs. NOOA scored 82.2% on SWE-bench Verified, 86.8% on CyberGym L1 tests using general-purpose agents with models such as GPT-5.5, and 50.2% on ARC-AGI-3 with GPT-5.5. Its ARC-AGI-3 score rose to 85.1% when paired with GPT-5.6-sol, with the cost per game below $20. Nvidia said NOOA’s memory design and removal of context-compaction processes effectively halved token usage compared with earlier designs, potentially lowering inference costs while improving output quality. The open-source release also supports Nvidia’s hardware business by making AI deployment cheaper and more accessible to a wider range of customers.